Hugging Face's logo

Datasets:
SLM-Lab
/
benchmark

Modalities:
Image
Formats:
imagefolder
Size:
1K - 10K
Libraries:
Datasets
Dataset card Data Studio Files
xet
Community
1
benchmark / docs /CROSSQ_TRACKER.md
kengzwl's picture
kengzwl
Upload folder using huggingface_hub
ebd7dfd verified
preview code
|
raw
history blame
20 kB

CrossQ Benchmark Tracker

Operational tracker for CrossQ benchmark runs. Updated by agent team.

Run Status

Wave 0 — Improvement Runs (COMPLETED, intake deferred)

Run Name Env Score (MA) Old Score Status Spec Name Intake
crossq-acrobot-v2 Acrobot-v1 -98.63 -108.18 ✅ solved crossq_acrobot ⬜ needs pull+plot
crossq-hopper-v8 Hopper-v5 1295.21 1158.89 ⚠️ improved crossq_hopper ⬜ needs pull+plot
crossq-swimmer-v7 Swimmer-v5 221.12 75.72 ✅ solved crossq_swimmer ⬜ needs pull+plot
crossq-invpend-v7 InvertedPendulum-v5 841.87 830.36 ⚠️ marginal crossq_inverted_pendulum ⬜ needs pull+plot
crossq-invdoubpend-v7 InvertedDoublePendulum-v5 4514.25 4952.63 ❌ worse, keep old crossq_inverted_double_pendulum ⬜ skip

Wave 1 — LayerNorm Experiments (COMPLETED)

Run Name Env Frames Score Spec Name Notes
crossq-humanoid-v2 Humanoid-v5 3M 2429.88 crossq_humanoid iter=4, 5.5h — VIOLATES 3h
crossq-hopper-ln-v2 Hopper-v5 3M 1076.76 crossq_hopper_ln LN +2% vs baseline
crossq-swimmer-ln-v2 Swimmer-v5 3M 22.90 crossq_swimmer_ln LN KILLED (-97%)
crossq-humanoid-ln-v2 Humanoid-v5 2M 506.65 crossq_humanoid_ln LN +19%, needs more frames

Wave 3 — Data Over Gradients (STOPPED — humanoid-ln-7m iter=1 inferior to iter=2)

Run Name Env Frames Score (at kill) Spec Name Notes
crossq-humanoid-ln-7m Humanoid-v5 7M 706 (at 70%) crossq_humanoid_ln Stopped — iter=2 reached 1850

Wave 2 — Full LN Sweep (RUNNING, just launched)

Run Name Env Frames Spec Name Notes
crossq-walker-ln Walker2d-v5 3M crossq_walker2d_ln 3890 — LN +22%! Near SAC 3900
crossq-halfcheetah-ln HalfCheetah-v5 3M crossq_halfcheetah_ln 6596 — LN -18% vs 8085
crossq-ant-ln Ant-v5 3M crossq_ant_ln 3706 — LN -5% vs 4046
crossq-invpend-ln InvertedPendulum-v5 3M crossq_inverted_pendulum_ln 731 — LN -13% vs 842
crossq-invdoubpend-ln InvertedDoublePendulum-v5 3M crossq_inverted_double_pendulum_ln 2727 — LN -45% vs 4953
crossq-cartpole-ln CartPole-v1 300K crossq_cartpole_ln 418 — LN +38%!
crossq-lunar-ln LunarLander-v3 300K crossq_lunar_ln 126 — LN -19% vs 136

Wave 4 — Extended-Frame LN (COMPLETED)

Run Name Env Frames Score (MA) Spec Name Notes
crossq-walker-ln-7m-v2 Walker2d-v5 7M 4277.15 crossq_walker2d_ln_7m ✅ BEATS SAC 3900! +10%
crossq-halfcheetah-ln-7m-v2 HalfCheetah-v5 6M 8784.55 crossq_halfcheetah_ln_7m +9% vs non-LN 8085, -10% SAC
crossq-ant-ln-7m-v2 Ant-v5 6M 5108.47 crossq_ant_ln_7m ✅ BEATS SAC 4844! +5%
crossq-hopper-ln-7m Hopper-v5 6M 1182 (at kill) crossq_hopper_ln_7m Stopped — LN hurts Hopper
crossq-walker-ln-i2 Walker2d-v5 3.5M 3766 (at kill) crossq_walker2d_ln_i2 Stopped — 7m run is better
crossq-invdoubpend-ln-7m InvertedDoublePendulum-v5 7M 5796 (at kill) crossq_inverted_double_pendulum_ln_7m Stopped — iter=2 much better

Wave 5 — iter=2 Gradient Density (COMPLETED)

Run Name Env Frames Score (MA) Spec Name Notes
crossq-humanoid-ln-i2-v2 Humanoid-v5 3.5M 1850.44 crossq_humanoid_ln_i2 +265% vs old 507! -29% SAC
crossq-invdoubpend-ln-i2-v2 InvertedDoublePendulum-v5 3.5M 7352.82 crossq_inverted_double_pendulum_ln_i2 +48% vs old 4953! -21% SAC

Wave 6 — WeightNorm Actor (COMPLETED)

Run Name Env Frames Score (MA) Spec Name Notes
crossq-humanoid-wn-v2 Humanoid-v5 7M 1681.45 crossq_humanoid_wn Strong but LN-i2 (1850) better
crossq-swimmer-wn-v2 Swimmer-v5 6M 165.49 crossq_swimmer_wn ❌ Regressed vs non-LN 221 (high variance)
crossq-hopper-wn Hopper-v5 6M 1097 (at kill) crossq_hopper_wn Stopped — not improving
crossq-walker-wn Walker2d-v5 7M 4124 (at kill) crossq_walker2d_wn Stopped — LN-7m better

Wave 7 — Next Improvement Runs (COMPLETED)

Run Name Env Frames Score (MA) Spec Name Notes
crossq-humanoidstandup-ln-i2 HumanoidStandup-v5 3.5M 150583.47 crossq_humanoid_standup_ln_i2 BEATS SAC 138222 (+9%)! LN + iter=2 + [1024,1024]
crossq-halfcheetah-ln-8m HalfCheetah-v5 7.5M 9969.18 crossq_halfcheetah_ln_8m BEATS SAC 9815 (+2%)! LN + iter=1, extended frames
crossq-hopper-i2 Hopper-v5 3.5M — crossq_hopper_i2 STOPPED — 101fps (9.6h), way over budget
crossq-invpend-7m InvertedPendulum-v5 7M — crossq_inverted_pendulum_7m Plain + iter=1, ~2.8h at 700fps

Wave 8 — v2 Final Runs (COMPLETED)

Run Name Env Frames Score (MA) Spec Name Notes
crossq-humanoidstandup-v2 HumanoidStandup-v5 2M 154162.28 crossq_humanoid_standup_v2 ✅ BEATS SAC +12%! LN iter=2, fewer frames
crossq-idp-v2 InvertedDoublePendulum-v5 2M 8255.82 crossq_inverted_double_pendulum_v2 ⚠️ Gap -9% vs SAC (was -21%). LN iter=2
crossq-walker-v2 Walker2d-v5 4M 4162.65 crossq_walker2d_v2 Near old 4277, beats SAC +33%. LN iter=1
crossq-humanoid-v2 Humanoid-v5 4M 1435.28 crossq_humanoid_v2 Below old 1850, high variance. LN iter=2
crossq-hopper-v2 Hopper-v5 3M 1150.08 crossq_hopper_v2 Below old 1295. iter=2 didn't help
crossq-ip-v3 InvertedPendulum-v5 3M 779.68 crossq_inverted_pendulum_v2 Below old 842. Seed variance
crossq-swimmer-v2 Swimmer-v5 3M 144.52 crossq_swimmer_v2 ❌ iter=2 disaster (was 221). Keep old

Scorecard — CrossQ vs SAC/PPO

Phase 1: Classic Control

Env CrossQ Best Other Gap LN Run?
CartPole-v1 418 (LN) 464 (SAC) -10% ✅ LN helps
Acrobot-v1 -98.63 -84.77 (SAC) close ✅ solved
LunarLander-v3 136.25 194 (PPO) -30% crossq-lunar-ln
Pendulum-v1 -163.52 -168 (SAC) ✅ beats done

Phase 2: Box2D

Env CrossQ Best Other Gap LN Run?
LunarLanderContinuous-v3 249.85 132 (PPO) ✅ beats done

Phase 3: MuJoCo

Env CrossQ Best Other Gap LN Run?
HalfCheetah-v5 9969 (LN 8M) 9815 (SAC) ✅ +2% BEATS SAC!
Hopper-v5 1295 1654 (PPO) -22% LN/WN both worse, keep baseline
Walker2d-v5 4277 (LN 7M) 3900 (SAC) ✅ +10% BEATS SAC!
Ant-v5 5108 (LN 6M) 4844 (SAC) ✅ +5% BEATS SAC!
Humanoid-v5 1850 (LN i2) 2601 (SAC) -29% Huge improvement from 507
HumanoidStandup-v5 154162 (LN i2 2M) 138222 (SAC) ✅ +12% BEATS SAC! v2
InvertedPendulum-v5 842 1000 (SAC) -16% LN hurts, keep baseline
InvertedDoublePendulum-v5 8256 (LN i2 2M) 9033 (SAC) -9% v2 improved from -21%
Reacher-v5 -5.66 -5.87 (SAC) ✅ beats done
Pusher-v5 -37.08 -38.41 (SAC) ✅ beats done
Swimmer-v5 221 301 (SAC) -27% WN regressed (165), keep baseline

Phase 4: Atari (PARKED — needs investigation before graduation)

Tested: Breakout, MsPacman, Pong, Qbert, Seaquest, SpaceInvaders

Status: Parked. Audit found issues — investigate CrossQ Atari performance before graduating. Atari CrossQ generally underperforms SAC. Investigate whether BRN warmup, lr tuning, or ConvNet-specific changes could help before publishing results.

Intake Checklist (per run)

  1. ⬜ Extract score: dstack logs NAME | grep trial_metrics → total_reward_ma
  2. ⬜ Find HF folder: huggingface_hub API query
  3. ⬜ Pull data: slm-lab pull SPEC_NAME
  4. ⬜ Update BENCHMARKS.md: score + HF link + status
  5. ⬜ Regenerate plot: slm-lab plot -t "ENV_NAME" -f FOLDER1,FOLDER2,...
  6. ⬜ Commit + push

Pending Fixes

  • Regenerate LunarLander plots with correct env name titles (564a6a96)
  • Universal env name audit across all plots (564a6a96)
  • Delete 58 stale Atari plots without -v5 suffix (564a6a96)
  • Wave 0 intake: pull HF data + regenerate plots (deferred — low bandwidth)

Decision Log

  • Swimmer-LN FAILED (22.90 final): LN hurts Swimmer. Non-LN 221.12 is best. Do NOT launch more Swimmer-LN runs.
  • Hopper-LN 3M (1076): WORSE than non-LN 6M (1295). More frames > LN for Hopper. Extended 6-7M LN run will tell if both helps.
  • LN HURTS most envs at 3M: HalfCheetah -18%, InvPend -13%, InvDoublePend -45%, Swimmer -97%. Only helps Humanoid (+19%).
  • Root cause: Critic BRN already normalizes. Actor LN over-regularizes, squashing activation scale on low/med-dim obs.
  • WeightNorm hypothesis: WN reparameterizes weights without squashing activations — should avoid LN's failure. Wave 6 testing.
  • Humanoid-v2 iter=4: MA 2923 at best session, likely beats SAC 2601. But uses iter=4 → ~150fps → 5.5h. VIOLATES 3h constraint. Not a valid CrossQ result.
  • Humanoid-LN 2M: 506.65. iter=1 is fast (700fps) but 2M not enough data. Launched 7M run (2.8h budget).
  • Frame budget rule: CrossQ at 700fps can do 7.5M in 3h. Use more frames than SAC, less than PPO.
  • InvDoublePend log_alpha_max=2.0: Failed (4514 vs old 4953). Default alpha cap better for this env.
  • CRITICAL: LN + extended frames REVERSES 3M findings — LN at 3M hurt most envs, but at 5-6M it BEATS non-LN baselines:
    • HalfCheetah-LN: -18% at 3M → +8% at 5M (8722 vs 8085). LN needs warmup frames.
    • Ant-LN: -5% at 3M → +25% at 5M (5054 vs 4046).
    • InvDoublePend-LN: -45% at 3M → +17% at 5M (5796 vs 4953).
    • Walker-LN: was already +22% at 3M, reached 4397 at 5.16M (74%) — beating SAC 3900.
  • iter=2 is the killer config for InvDoublePend: 7411 at 69% completion, 50% above baseline, approaching SAC 9359.
  • WN promising: Swimmer-WN 255 > non-LN 221. Walker-WN 4124 strong. Need full runs to confirm.
  • RunPod batch eviction: All 13 runs killed at 01:25 UTC. Root cause: dstack credits depleted.
  • Strategic triage: After relaunch, stopped 6 redundant/underperforming runs, kept 7 promising:
    • KEPT: walker-ln-7m (beating SAC), ant-ln-7m (beating SAC), halfcheetah-ln-7m (closing gap), invdoubpend-ln-i2 (iter=2 best), swimmer-wn (WN solving), humanoid-ln-i2 (best Humanoid), humanoid-wn (alternative)
    • STOPPED: hopper-ln-7m (LN hurts), hopper-wn (flat), walker-ln-i2 (7m better), walker-wn (7m better), invdoubpend-ln-7m (i2 much better), humanoid-ln-7m (i2 better)
  • FINAL RESULTS (7 runs completed):
    • Walker-LN-7m: 4277 — BEATS SAC 3900 (+10%)
    • Ant-LN-7m: 5108 — BEATS SAC 4844 (+5%)
    • HalfCheetah-LN-7m: 8785 — gap narrowed from -17% to -10%
    • InvDoublePend-LN-i2: 7353 — gap narrowed from -47% to -21%
    • Humanoid-LN-i2: 1850 — massive improvement from 507 (-29% vs SAC)
    • Humanoid-WN: 1681 — strong but LN-i2 wins
    • Swimmer-WN: 165 — REGRESSED from 221 (high variance, consistency=-0.79). WN does NOT fix Swimmer.
  • LN + extended frames confirmed: The universal recipe is LN actor + more frames. Works for 5/7 MuJoCo envs. Exceptions: Hopper (LN hurts regardless), Swimmer (LN kills, WN also fails at full run).
  • Swimmer paradox: WN looked promising at 67% (MA 255) but regressed to 165 at completion. High session variance. Non-LN 221 remains best.
  • Humanoid strategy: LN+iter=2 (1850) > WN (1681) > LN+iter=1 7M (706). Humanoid needs gradient density, not just data.
  • Hopper-i2 too slow: 101fps with iter=2 [512,512], would take 9.6h. Stopped. Plain baseline at 1295 with 5M/iter=1 (700fps) is best. Hopper is CrossQ's weakest MuJoCo env — 22% below PPO 1654, no normalization variant helps.
  • Wave 7 launched: HumanoidStandup-LN-i2 (353fps, early MA 106870 vs baseline 115730), HalfCheetah-LN-8m (708fps), InvPend-7m (plain, more data).

Atari Investigation

Current CrossQ vs SAC Atari Scores

Game CrossQ SAC Ratio Verdict
Breakout 0.91 20.23 4.5% catastrophic
MsPacman 238.51 1336.96 17.8% catastrophic
Pong -20.82 10.89 no learning catastrophic
Qbert 4268.66 3331.98 128% CrossQ wins
Seaquest 216.19 1565.44 13.8% catastrophic
SpaceInvaders 360.37 507.33 71% poor

CrossQ wins 1/6 games (Qbert). The other 5 show near-total failure, with 3 games at <18% of SAC performance.

Root Cause Analysis

Primary hypothesis: BRN placement is wrong for ConvNets.

The CrossQ Atari critic architecture places a single LazyBatchRenorm1d layer after the final FC layer (post-Flatten, post-Linear(512)). This is fundamentally different from the MuJoCo architecture where BRN layers are placed between every hidden FC layer (two BRN layers for [256,256], two for [512,512], etc.).

Atari critic (1 BRN layer):

Conv2d(32) -> ReLU -> Conv2d(64) -> ReLU -> Conv2d(64) -> ReLU -> Flatten -> Linear(512) -> BRN -> ReLU

MuJoCo critic (2 BRN layers):

Linear(W) -> BRN -> ReLU -> Linear(W) -> BRN -> ReLU

The CrossQ paper's core insight is that BN/BRN statistics sharing between current and next-state batches replaces target networks. With only one BRN layer after 512-dim features, the normalization may be insufficient — the ConvNet backbone (3 conv layers) processes current and next-state images with NO shared normalization. The BRN only operates on the final FC representation. This means the cross-batch statistics sharing that eliminates the need for target networks is weak.

Secondary hypothesis: Hyperparameters ported directly from MuJoCo without ConvNet adaptation.

Key differences between CrossQ Atari vs SAC Atari specs:

Parameter CrossQ Atari SAC Atari Issue
lr 1e-3 3e-4 3.3x higher — too aggressive for ConvNets
optimizer Adam AdamW No weight decay in CrossQ
betas [0.5, 0.999] [0.9, 0.999] Low beta1 for ConvNets is risky
clip_grad_val 0.5 0.5 same
loss SmoothL1Loss SmoothL1Loss same
policy_delay 3 1 (default) Delays policy updates 3x
log_alpha_max 0.5 none (uses clamp [-5, 2]) Tighter alpha cap
warmup_steps 10000 n/a Only 10K for Atari
target networks none polyak 0.005 CrossQ core difference
init_fn orthogonal_ orthogonal_ same

The lr=1e-3 with betas=[0.5, 0.999] combination is specifically tuned for MuJoCo MLPs per the CrossQ paper. ConvNets are known to be more sensitive to learning rates — SAC Atari uses lr=3e-4 which is standard for Atari. The low beta1=0.5 reduces momentum, which may cause unstable gradient updates in ConvNets where feature maps evolve slowly.

Tertiary hypothesis: BRN warmup_steps=10000 is too low for Atari.

At training_frequency=4 and num_envs=16, each training step consumes 64 frames. With training_iter=3, there are 3 gradient steps per training step. So 10K warmup means 10K BRN steps = 10K/3 = ~3333 training steps = ~213K frames (10.7% of 2M). During warmup, BRN behaves as standard BN (r_max=1, d_max=0), which has been shown to cause divergence in RL (see CrossQ standard BN results in MEMORY.md).

MuJoCo uses warmup_steps=100000 = 100K BRN steps. At training_frequency=1 and num_envs=16, that's ~1.6M frames (significant fraction of typical 3-7M runs). This much slower warmup gives the running statistics time to stabilize. Atari at 10K warmup transitions to full BRN correction far too early when running statistics are still poor.

Fourth hypothesis: Cross-batch forward is ineffective for ConvNets.

In calc_q_cross_discrete, states and next_states are concatenated and passed through the critic together. For MuJoCo (small state vectors), this is cheap and effective — BN statistics computed over both batches provide good normalization. For Atari (84x84x4 images), the concatenated batch goes through 3 conv layers with NO normalization, then hits a single BRN layer at dim=512. The conv layers see a batch that mixes current and next frames, but without BN in the conv layers, this mixing provides no cross-batch regularization benefit. The entire CrossQ mechanism reduces to "BRN on the last FC layer of a frozen ConvNet backbone."

Proposed Fixes (Priority Order)

P0: Lower learning rate to SAC-Atari defaults

  • Change lr: 1e-3 to lr: 3e-4 for both actor and critic
  • Change betas: [0.5, 0.999] to default [0.9, 0.999]
  • Rationale: The lr=1e-3/beta1=0.5 combo is CrossQ-paper MuJoCo-specific. ConvNets need conservative lr.

P1: Increase BRN warmup to 100K steps

  • Change warmup_steps: 10000 to warmup_steps: 100000
  • Rationale: Match MuJoCo proportionally. 100K BRN steps at iter=3 = ~2.1M frames, which is the full run. This means BRN stays in near-standard-BN mode for most of training — essentially disabling the full BRN correction that may be destabilizing ConvNets.

P2: Add BRN after each conv layer (deeper cross-batch normalization)

  • Place LazyBatchRenorm1d (or BatchRenorm2d, which would need implementation) after each Conv2d layer
  • Rationale: The CrossQ paper's mechanism relies on shared BN statistics between current/next batches. With BRN only at the FC layer, the ConvNet backbone has no cross-batch normalization, defeating the purpose.
  • Note: This requires implementing BatchRenorm2d (2D spatial variant). Standard BatchNorm2d normalizes per-channel across spatial dims — a BatchRenorm2d would do the same with correction factors.
  • Risk: This is a code change, not a spec-only fix. Higher effort.

P3: Remove policy_delay for Atari

  • Change policy_delay: 3 to policy_delay: 1
  • Rationale: SAC Atari uses no policy delay. With only 2M frames and iter=3, policy_delay=3 means the policy is updated once every 3 critic updates. Combined with the already-low frame budget, the policy may not get enough gradient updates to learn.
  • Total policy updates at 2M frames: (2M / (4 * 16)) * 3 / 3 = 31,250. Without delay: 93,750. 3x more policy updates.

P4: Switch to AdamW with weight decay

  • Match SAC Atari's AdamW with eps: 0.0001
  • Rationale: Weight decay provides implicit regularization that may partially compensate for the missing target network smoothing.

Experiment Plan

  1. Exp A (spec-only, highest impact): lr=3e-4, betas=[0.9,0.999], warmup=100K, policy_delay=1. Test on Pong + Breakout (fast signal games).
  2. Exp B (spec-only): Same as A but keep policy_delay=3. Isolates lr/warmup effect.
  3. Exp C (spec-only): Same as A but lr=1e-3 (keep CrossQ lr). Isolates beta/warmup effect.
  4. Exp D (code change): Add BatchRenorm2d after conv layers. Test with Exp A settings.

If Exp A solves the problem, no code changes needed. If not, Exp D addresses the fundamental architectural mismatch.

Key Insight

The Qbert success is telling. Qbert has relatively simple visual patterns and discrete state changes — the ConvNet can extract good features even with aggressive lr. Games like Pong and Breakout require precise spatial reasoning where ConvNet feature quality matters more, and the aggressive lr/low-momentum combo destabilizes learning before features mature.